Kahan summation is basically just synthesizing a larger floating point type (with 46/106 bits of mantissa for 32/64-bit base type) and then using that to do the

Kahan summation is basically just synthesizing a larger floating point type (with 46/106 bits of mantissa for 32/64-bit base type) and then using that to do the sum, though. You still lose precision if your errors exceed that new, larger margin. Eg try adding 34000 + 56 + .078 - 56 - 34000, carrying two digits per float.